Refactor generation backend

[StAlKeR7779]

What type of PR is this? (check all applicable)

• Refactor
• Feature
• Bug Fix
• Optimization
• Documentation Update
• Community Node Submission

Have you discussed this change with the InvokeAI team?

• Yes
• No, because:

Have you updated all relevant documentation?

• Yes
• No

Description

• Remove SDXL raw prompt nodes
• SDXL and SD1/2 generation merged to same nodes - t2l/l2l
• Fixed - if no xformers installed we trying to enable attention slicing, ignoring torch-sdp availability
• Fixed - In SDXL negative prompt now creating zeroed tensor(according to official code)
• Added mask field to l2l node
• Removed inpaint node and all legacy code related to this node
• Pass info about seed in latents, so we can use it to initialize ancestral/sde schedulers
• t2l and l2l nodes moved from strength to denoising_start/end
• Removed code for noise threshold(@hipsterusername said that there no plans to restore this feature)
• Fixed - first preview image now not gray
• Fixed - report correct total step count in progress, added scheduler order in progress event
• Added MaskEdge and ColorCorrect nodes (@hipsterusername)

Added/updated tests?

• Yes
• No

[keturn]

Can we merge #4175 first?

I admit it's partly my thirst for those sweet line-deletion credits, but also there are a half-dozen different things going on in a single PR here and if we can take advantage of an already-existing PR to cut it down in scope, I think that will make it easier to review.

[psychedelicious]

🔥

[keturn]

Thanks for merging in the other code removal first!

That knocked down the size of this PR a bit, but IMHO it's still too much for a single PR.

There's inpainting stuff that could be done without the SDXL stuff,
the seeding of the stochastic schedulers is an independent issue,
the changes to adjust-memory-efficient-attention could and should be considered and tested on their own--

I've left a few comments so far, but this does not constitute anything near a complete review. I haven't even got to the changes in shared_invokeai_diffusion yet, let alone any of the SDXL.